Re: Strage values from FIBMAP on holes or after end of file

[Steve Lord <lord@sgi.com>]

> 
> > p.s. I have an xfs filesystem which has existed for probably a year or so
> > now, it is where I do all my kernel builds, merges with linus's trees, rpm
> > building etc. I also frequently have multiple operations on going in parall
> el.
> > It currently looks like this:
> > 
> > df /src
> > Filesystem           1k-blocks      Used Available Use% Mounted on
> > /dev/hda5              4194932   2439204   1755728  59% /src
> > 
> > df -i /src
> > Filesystem            Inodes   IUsed   IFree IUse% Mounted on
> > /dev/hda5            4200896  113588 4087308    3% /src
> > 
> > xfs_db: frag
> > actual 109889, ideal 109553, fragmentation factor 0.31%
> > xfs_db: freesp
> >    from      to extents  blocks    pct
> >       1       1     597     597   0.14
> >       2       3     414     967   0.22
> >       4       7     333    1762   0.40
> >       8      15     279    3165   0.72
> >      16      31     153    3458   0.79
> >      32      63     148    6904   1.57
> >      64     127     259   24360   5.55
> >     128     255     290   52605  11.98
> >     256     511     186   65816  14.99
> >     512    1023      71   49383  11.25
> >    1024    2047      14   19025   4.33
> >    2048    4095       6   17071   3.89
> >    4096    8191       8   50298  11.46
> >    8192   16383       4   49448  11.27
> >   16384   32767       1   21222   4.83
> >   32768   65535       2   72851  16.60
> > 
> > ./fibmap /src
> > Result for 106819 files 6584 dirs: Int: 0.891066 Ext: 0.999941 0.999942 0.9
> 99954 0.999954 - 1.347378 - Path 1.495712 2.525242
> > Mean Internal frag: 0.891066
> > Mean External frag: 0.999941
> > Mean Frag. Path 1.495712
> I cant say much, as I dont know how xfs_db calculates these values.
> 
> But I am wondering about the internal frag. Thought xfs only can handel
> 4K blocks?
> If your /src fs uses 4K blocks, then 0.89 e.g. 10 % is a very low value
> of internal fragmentation.

It is using 4K blocks, XFS is good at this stuff ;-). Actually, large
files will tend to reduce the fragmentation, i.e. a 399K file which
consumes 400K of disk space is not very fragmented by this measure,
so really this is a function of file size distribution.

> If it uses 1K (as shown in the ooutput of df) then it is quite normal
> compared to a 1K ext2 system.

Blocksize in df output is not a function of filesystem blocksize, it is
always reported in 1K chunks. Also be careful of st_blksize reported by
stat, it is the optimum size for doing I/O, not the filesystem block
size (not that you are using this right now).

Steve


> 
> Constantin